sonora
Starting the speech engine

Settings

Recognition model

Both models recognise English and German and tell speakers apart. The selected one is used for every recording and file.

High quality

A second, stronger speech model (NVIDIA Parakeet) reads every finished stretch a few seconds behind the live transcript and replaces the words, with far fewer errors in names, numbers, and punctuation. Files get one pass over the whole recording. Speaker labels and the instant text still come from the live model.

Parakeet TDT 0.6B v3English, German, and other European languages. Names and terms you type are not passed to it; the live model still spells them.

Parakeet (CC-BY-4.0, NVIDIA) downloads once from huggingface.co, or from your organisation’s mirror, and stays in this browser.

Summaries and clean-up

A small language model that writes summaries and polishes dictation, on this computer, after names and contact details have been removed. Transcription works without it.

Qwen3 (Apache-2.0) downloads once from huggingface.co, or from your organisation’s mirror, and stays in this browser.

Keyboard shortcuts

R
Start or stop recording
Space
Play or pause the recording
← →
Skip back or forward 5 seconds
O
Open an audio file
N
New session
P
Privacy mode on or off
Ctrl K
Search the transcript
Ctrl S
Export the transcript in the last used format
Ctrl Shift C
Copy the transcript
Esc
Close search, dialogs, and editors
?
Show this list

Click a timestamp to listen from there. Click a speaker name to rename it everywhere. Click any line of text to correct it.

Leave privacy mode?

This session exists only in memory. Save it to the library, or wipe it.

Privacy mode

For conversations that must leave no trace.